- Created emitters for jitted load and store and apply to MVN node
- MVN template function removed
- Optimized tiles processing in MVN node implementation
* Move Convolution and ConvolutionBackpropData ref impls into separate files.
* Add convolution unit tests.
* New convolution reference implementation.
* Remove unused convolution ref impl argument.
* Fix style.
* Revert "Remove unused convolution ref impl argument."
This reverts commit 739065d0d0.
* WA for arm-plugin: additional include with ConvolutionBackpropData.
* Style format in Convolution SLT CPU instantiation.
* Add 1D Convolution SLT CPU tests.
* Add Convolution Serialization SLT.
* Update source banners with 2021 date.
* Specification review.
* Readability improvement in padding detection.
* Refactoring regarding Tensor usage.
* Iteration over tensor slices made more readable.
* Code refactored to use only one convolution implementation.
3D convolution is used to compute also in 1D & 2D case (parameters,
inputs and filters shapes are adjusted accordingly).
* Removed Tensor abstraction.
* Name unnamed namespace as convolution_details.
* Refactoring: replaced std::next + negative index with std::prev.
* Specification refactoring.
* Revert "Name unnamed namespace as convolution_details."
This reverts commit cea526ec49.
* Added new convolution() overload.
* Fix legacy convolution() overload (needed for kmb-plugin).
* Reduced number of template type arguments in convolution ref impl.
* Added 'output' section in Convolution spec.
* Remove floating round type configuration.
* eliminate Unsqueeze+Gather pair, when Gather gathers data by 1 dimension which was previously added by Unsqueeze which is actually doing nothing.
* calculate K only once in StaticShapeTopK. The problem happens when we have ShapeOf->Concat->ReduceMin subgraph for K evaluation. If we have a pretty small input size, the value that we received from ShapeOf may be less than one that it is concatenated with (e.g. ShapeOf 283 vs const 300), so ReduceMin returns 283. After ShapeOf elimination we don't have a chance to propagate 283 so we get 300 as a result and shape inference fail then. There are no problems with bigger input sizes just because ShapeOf always propagates value >300 and there are no such mismatch.
* ConvertLike: Develop reference implementation
* ConvertLike: Enable single layer tests for GPU plugin
* ConvertLike: Enable bf16 precision for evaluate method
* ConvertLike: Add unit tests
* ConvertLike: Add dynamic shape test case
* ConvertLike: Remove unnecessary ngraph namespace and using declaration for v1::ConvertLike
* ConvertLike: Simplified reference::convert by using std::enable_if
* fix initialization bug of spatial_scale in tests (affected input generating)
* fix input generating for bilinear ROI Pooling
* correct parameters for myriad tests:
* myriad plugin does not support batch for this layer;
* decrease threshold since myriad uses fp16 calculations
* Support DTS for GatherElements
* Extract GatherBase to a common part
* Introduce tests on inference
* Introduce tests on function comparing
* Disable failing tests
Before this patch constant with weights could be not detected if
it wasn't directly connected to Conv/Deconv layer.
Now weights always uses common data format (bfzyx) in the plugin which is
converted into weights format later (goiyx, oiyx, etc), so weights sub-graph
can now contain anything
Anton Pankratv